How difficult would it be to convert a synthesis by rule program for English (DECtalk) to speak Japanese? What is the ideal, hopefully minimal, corpus of recorded materials that would have to be spectrally analyzed in order to derive sufficient information for a first order approximation to Japanese? We have recorded three male speakers reading a list of a little over 100 nonsense syllables, which exhaust the CV(C) syllable inventory of Japanese, and provide information about segmental contrasts. One of the speakers then read a 500-word text to provide information on sentence-level phenomena. We have measured segmental durations, formant motions, and deletion phenomena in broadband spectrograms of this recording, and are beginning to formulate rules for synthesis on the basis of computer analyses of the data. The approach now appears far less unreasonable than it did when we began. Some results of the analysis will be presented. [Work supported in part by an NIH grant.]